❄Snowflake | Cortex Agents & dbt

Opening statement

Beyond a Reasonable dbt

Trustworthy pipelines for AI agents

Executive briefing Managing the Cortex Agent lifecycle as code

The docket

What we'll cover

01The shift: agents are the new front door to data
02The risk: why most agents can't be trusted
03The reframe: treat agents like production software
04How it works: one project, the whole lifecycle
05The payoff: measurable trust, governed delivery

The shift

Every team wants an AI agent

Conversational agents are becoming the front door to enterprise data. Business users ask questions in plain language and expect a trustworthy answer, without writing SQL or waiting on a report.

The demand is real and accelerating. The question is no longer whether you'll build agents; it's whether you can trust the ones you ship.

The case

But most agents are built by click-ops

The charge

Assembled by hand in a UI
No version control or history
No peer review before production
No tests: accuracy is a vibe
Dev and prod quietly drift apart

The consequence

Wrong answers, no way to diff or roll back
One person holds the whole thing
Every environment rebuilt by hand
Trust erodes on the first bad answer
Adoption stalls

The reframe

Treat AI agents like production software

An agent is only as trustworthy as the process behind it. So manage the entire lifecycle as code (the data model, the agent, its tests, and its schedule) natively on Snowflake, with dbt.

Same discipline you already apply to the rest of your data platform: version it, review it, test it, promote it.

The payoff

What "as code" delivers

1

Versioned & reviewed

Every change is in Git, diffed, and approved in a pull request before it reaches users.

2

Measurable trust

Evaluations score the agent against known answers; ship only when it clears the bar.

3

One codebase

The same code targets dev, staging, and prod. No rebuilding by hand per environment.

4

Ship anywhere

Deliver the same governed agent to Snowflake Intelligence, Teams, an API, or MCP.

The evidence

One project builds all of it

One dbt project on Snowflake builds every box, and the loop is where evaluation scores drive the next improvement.

The verdict

≥95%

Trust becomes a number

Before an agent ships, it's scored against a set of known questions and answers. The bar to promote: at least 95% answer correctness. Fall short, fix the right layer, and re-run.

Accuracy stops being a gut feeling and becomes a metric you can track, compare across versions, and hold the line on.

Closing arguments

Ship it where people already work

💬

Snowflake Intelligence

Ask-your-data experience

👥

Microsoft Teams

In the flow of work

🔌

REST API

Embed in your apps

🧩

MCP

Open agent interop

One governed agent, delivered to every surface your users already live in.

The ruling

The bottom line

✓AI agents are only as trustworthy as the process behind them.
✓Manage the whole lifecycle as code: versioned, reviewed, tested, promoted.
✓Trust becomes measurable; delivery becomes governed and repeatable.